Photo to STL: a relief workflow, not a magic 3D scanner

A single image gives the model perspective, occlusion order, relative size, texture gradients, and lighting cues. It does not give the hidden side of a face, the inside of a cup, or the space behind a chair. It also cannot separate a hard cast shadow from a dark surface, because both arrive as a dark region next to the object that cast it. Those limits are not a defect in the model; they are missing information.

The honest output is a front-facing relief. Treat it as a prepared manufacturing input, not a measured reconstruction. If the project needs a full 3D object, use multiple views, a scan, or a sculpt. If it needs a plaque, a lithophane, or a carved panel, a relief is the right shape of answer.

Use the sharpest original you have. Crop to the subject and leave a quiet border that can become a base. Front-facing portraits, pets, products, and architectural details work better than groups, glass, mirrors, smoke, and fine hair against a busy background. Avoid heavy filters, HDR halos, and aggressive sharpening; those become ridges in the depth field.

If the subject is small in the frame, crop before generating. If the background is busy, flatten it or remove it. If the image is a JPEG, inspect it at 100% for compression blocks around edges. A clean, well-lit phone photo is usually better than an old, heavily compressed studio file.

The first free preview answers whether the scene order is correct: is the face in front of the wall, is the pet in front of the sofa, is the product standing on the table? If the order is wrong, changing the mesh or the material will not fix it. Go back to the source or choose a different image.

Once the order is right, remap the range. Raise the black point to quiet the background, lower the white point if a small near region is compressing the face, and use contrast to separate the subject. Invert if the receiving software expects the opposite direction. Keep the source visible while adjusting so a soft error does not become a sharp ridge.

A portrait plaque, a lithophane, and a carved furniture panel have different physical constraints. Decide the finished width first, then the relief height and base. A shallow relief reads well at small sizes and keeps the cutter or nozzle out of trouble. A deep relief can look dramatic but may create undercuts, unsupported overhangs, or tool collisions.

The manufacturing checks compare the selected detail diameter, depth step, stock thickness, tonal slices, slope, base, and bounding box. They are starting points, not a machine certificate. Confirm the real tool, material, and fixture in CAM or the slicer before cutting or printing.

The free path includes three AI previews per day, the shaping controls, a 3D preview, an 8-bit PNG, a compatibility 16-bit PNG, and a low-resolution preview STL. Use the preview STL to check scale and orientation in the target program. It is deliberately coarse, so do not judge fine detail from it.

One approved production job costs one credit and returns raw float EXR, verified true 16-bit forward and inverse PNGs, a denser watertight STL, a preview, and a manifest. New accounts start with one credit. A confirmed provider failure returns the debit; a completed result you decide not to use does not.

For 3D printing, slice a small version at the final layer height and inspect the first layers, overhangs, and thin features. For CNC, simulate the toolpath and cut a scrap sample from the same material batch. A test that changes the material, tool, or orientation proves much less than a test that reproduces the production setup.

Keep the source photo, the STL, and the manifest together. If the customer asks for a larger version or a deeper relief, the saved decisions explain how to reproduce the result instead of starting from memory.